Deciphering the Temporal Evolution of Key Controlling Factors in Shale Oil Production Using Interpretable Machine Learning

Precise forecasting of shale oil production is essential for optimizing reservoir development and guiding field-scale operational decisions. This study presents an integrated machine learning framework combining multistage cumulative production prediction, robust feature selection, and model interpretation. Specifically, a cascaded prediction architecture is employed to capture temporal dependencies: it leverages early-stage production as auxiliary inputs to improve forecasts for 90-, 180-, and 360 day cumulative production. To ensure model stability under limited data conditions, multiple regression algorithms (Ridge, CatBoost, Random Forest, and Lasso) are benchmarked across seven production horizons. Meanwhile, an ensemble-based consensus feature selection strategy is applied to identify a compact set of temporally stable variables. The proposed cascaded modeling strategy delivers substantial performance gains, improving predictive accuracy by approximately 30% in R2 (from 0.45 to 0.75) and reducing RMSE by 28% relative to independent single-stage models, demonstrating enhanced robustness. Model interpretation using SHapley additive explanations and partial dependence plots reveals that early production is primarily controlled by operational parameters, while geological characteristics increasingly govern long-term output. Nonlinear effects and key feature interactions reflect physically meaningful mechanisms, including diminishing returns from increased fracture length and pressure-sensitive production behavior. Overall, the proposed workflow provides a reusable and field-adaptable predictive framework, offering practical guidance for optimizing the completion design and production management in shale oil reservoirs.